/* //////////////////////////////////////////////////////////////////////////// -- Testing sgeqrf */ int main( int argc, char** argv) { TESTING_INIT(); const float d_neg_one = MAGMA_D_NEG_ONE; const float d_one = MAGMA_D_ONE; const float c_neg_one = MAGMA_S_NEG_ONE; const float c_one = MAGMA_S_ONE; const float c_zero = MAGMA_S_ZERO; const magma_int_t ione = 1; real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0; float Anorm, error=0, error2=0; float *h_A, *h_R, *tau, *h_work, tmp[1]; magmaFloat_ptr d_A, dT; magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb, size; magma_int_t ISEED[4] = {0,0,0,1}; magma_opts opts; opts.parse_opts( argc, argv ); magma_int_t status = 0; float tol = opts.tolerance * lapackf77_slamch("E"); // version 3 can do either check if (opts.check == 1 && opts.version == 1) { opts.check = 2; printf( "%% version 1 requires check 2 (solve A*x=b)\n" ); } if (opts.check == 2 && opts.version == 2) { opts.check = 1; printf( "%% version 2 requires check 1 (R - Q^H*A)\n" ); } printf( "%% version %d\n", (int) opts.version ); if ( opts.check == 1 ) { printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |R - Q^H*A| |I - Q^H*Q|\n"); printf("%%==============================================================================\n"); } else { printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |b - A*x|\n"); printf("%%===============================================================\n"); } for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[itest]; N = opts.nsize[itest]; min_mn = min( M, N ); lda = M; n2 = lda*N; ldda = magma_roundup( M, opts.align ); // multiple of 32 by default nb = magma_get_sgeqrf_nb( M, N ); gflops = FLOPS_SGEQRF( M, N ) / 1e9; // query for workspace size lwork = -1; lapackf77_sgeqrf( &M, &N, NULL, &M, NULL, tmp, &lwork, &info ); lwork = (magma_int_t)MAGMA_S_REAL( tmp[0] ); TESTING_MALLOC_CPU( tau, float, min_mn ); TESTING_MALLOC_CPU( h_A, float, n2 ); TESTING_MALLOC_CPU( h_work, float, lwork ); TESTING_MALLOC_PIN( h_R, float, n2 ); TESTING_MALLOC_DEV( d_A, float, ldda*N ); if ( opts.version == 1 || opts.version == 3 ) { size = (2*min(M, N) + magma_roundup( N, 32 ) )*nb; TESTING_MALLOC_DEV( dT, float, size ); magmablas_slaset( MagmaFull, size, 1, c_zero, c_zero, dT, size ); } /* Initialize the matrix */ lapackf77_slarnv( &ione, ISEED, &n2, h_A ); lapackf77_slacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda ); magma_ssetmatrix( M, N, h_R, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ nb = magma_get_sgeqrf_nb( M, N ); gpu_time = magma_wtime(); if ( opts.version == 1 ) { // stores dT, V blocks have zeros, R blocks inverted & stored in dT magma_sgeqrf_gpu( M, N, d_A, ldda, tau, dT, &info ); } else if ( opts.version == 2 ) { // LAPACK complaint arguments magma_sgeqrf2_gpu( M, N, d_A, ldda, tau, &info ); } #ifdef HAVE_CUBLAS else if ( opts.version == 3 ) { // stores dT, V blocks have zeros, R blocks stored in dT magma_sgeqrf3_gpu( M, N, d_A, ldda, tau, dT, &info ); } #endif else { printf( "Unknown version %d\n", (int) opts.version ); return -1; } gpu_time = magma_wtime() - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) { printf("magma_sgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); } if ( opts.check == 1 && (opts.version == 2 || opts.version == 3) ) { if ( opts.version == 3 ) { // copy diagonal blocks of R back to A for( int i=0; i < min_mn-nb; i += nb ) { magma_int_t ib = min( min_mn-i, nb ); magmablas_slacpy( MagmaUpper, ib, ib, &dT[min_mn*nb + i*nb], nb, &d_A[ i + i*ldda ], ldda ); } } /* ===================================================================== Check the result, following zqrt01 except using the reduced Q. This works for any M,N (square, tall, wide). Only for version 2, which has LAPACK complaint output. Or for version 3, after restoring diagonal blocks of A above. =================================================================== */ magma_sgetmatrix( M, N, d_A, ldda, h_R, lda ); magma_int_t ldq = M; magma_int_t ldr = min_mn; float *Q, *R; float *work; TESTING_MALLOC_CPU( Q, float, ldq*min_mn ); // M by K TESTING_MALLOC_CPU( R, float, ldr*N ); // K by N TESTING_MALLOC_CPU( work, float, min_mn ); // generate M by K matrix Q, where K = min(M,N) lapackf77_slacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq ); lapackf77_sorgqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info ); assert( info == 0 ); // copy K by N matrix R lapackf77_slaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr ); lapackf77_slacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr ); // error = || R - Q^H*A || / (N * ||A||) blasf77_sgemm( "Conj", "NoTrans", &min_mn, &N, &M, &c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr ); Anorm = lapackf77_slange( "1", &M, &N, h_A, &lda, work ); error = lapackf77_slange( "1", &min_mn, &N, R, &ldr, work ); if ( N > 0 && Anorm > 0 ) error /= (N*Anorm); // set R = I (K by K identity), then R = I - Q^H*Q // error = || I - Q^H*Q || / N lapackf77_slaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr ); blasf77_ssyrk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr ); error2 = safe_lapackf77_slansy( "1", "Upper", &min_mn, R, &ldr, work ); if ( N > 0 ) error2 /= N; TESTING_FREE_CPU( Q ); Q = NULL; TESTING_FREE_CPU( R ); R = NULL; TESTING_FREE_CPU( work ); work = NULL; } else if ( opts.check == 2 && M >= N && (opts.version == 1 || opts.version == 3) ) { /* ===================================================================== Check the result by solving consistent linear system, A*x = b. Only for versions 1 & 3 with M >= N. =================================================================== */ magma_int_t lwork2; float *x, *b, *hwork; magmaFloat_ptr d_B; // initialize RHS, b = A*random TESTING_MALLOC_CPU( x, float, N ); TESTING_MALLOC_CPU( b, float, M ); lapackf77_slarnv( &ione, ISEED, &N, x ); blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_zero, b, &ione ); // copy to GPU TESTING_MALLOC_DEV( d_B, float, M ); magma_ssetvector( M, b, 1, d_B, 1 ); if ( opts.version == 1 ) { // allocate hwork magma_sgeqrs_gpu( M, N, 1, d_A, ldda, tau, dT, d_B, M, tmp, -1, &info ); lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] ); TESTING_MALLOC_CPU( hwork, float, lwork2 ); // solve linear system magma_sgeqrs_gpu( M, N, 1, d_A, ldda, tau, dT, d_B, M, hwork, lwork2, &info ); if (info != 0) { printf("magma_sgeqrs returned error %d: %s.\n", (int) info, magma_strerror( info )); } TESTING_FREE_CPU( hwork ); } #ifdef HAVE_CUBLAS else if ( opts.version == 3 ) { // allocate hwork magma_sgeqrs3_gpu( M, N, 1, d_A, ldda, tau, dT, d_B, M, tmp, -1, &info ); lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] ); TESTING_MALLOC_CPU( hwork, float, lwork2 ); // solve linear system magma_sgeqrs3_gpu( M, N, 1, d_A, ldda, tau, dT, d_B, M, hwork, lwork2, &info ); if (info != 0) { printf("magma_sgeqrs3 returned error %d: %s.\n", (int) info, magma_strerror( info )); } TESTING_FREE_CPU( hwork ); } #endif else { printf( "Unknown version %d\n", (int) opts.version ); return -1; } magma_sgetvector( N, d_B, 1, x, 1 ); // compute r = Ax - b, saved in b blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_neg_one, b, &ione ); // compute residual |Ax - b| / (max(m,n)*|A|*|x|) float norm_x, norm_A, norm_r, work[1]; norm_A = lapackf77_slange( "F", &M, &N, h_A, &lda, work ); norm_r = lapackf77_slange( "F", &M, &ione, b, &M, work ); norm_x = lapackf77_slange( "F", &N, &ione, x, &N, work ); TESTING_FREE_CPU( x ); TESTING_FREE_CPU( b ); TESTING_FREE_DEV( d_B ); error = norm_r / (max(M,N) * norm_A * norm_x); } /* ===================================================================== Performs operation using LAPACK =================================================================== */ if ( opts.lapack ) { cpu_time = magma_wtime(); lapackf77_sgeqrf( &M, &N, h_A, &lda, tau, h_work, &lwork, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) { printf("lapackf77_sgeqrf returned error %d: %s.\n", (int) info, magma_strerror( info )); } } /* ===================================================================== Print performance and error. =================================================================== */ printf("%5d %5d ", (int) M, (int) N ); if ( opts.lapack ) { printf( "%7.2f (%7.2f)", cpu_perf, cpu_time ); } else { printf(" --- ( --- )" ); } printf( " %7.2f (%7.2f) ", gpu_perf, gpu_time ); if ( opts.check == 1 ) { bool okay = (error < tol && error2 < tol); status += ! okay; printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") ); } else if ( opts.check == 2 ) { if ( M >= N ) { bool okay = (error < tol); status += ! okay; printf( "%10.2e %s\n", error, (okay ? "ok" : "failed") ); } else { printf( "(error check only for M >= N)\n" ); } } else { printf( " ---\n" ); } TESTING_FREE_CPU( tau ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_work ); TESTING_FREE_PIN( h_R ); TESTING_FREE_DEV( d_A ); if ( opts.version == 1 || opts.version == 3 ) { TESTING_FREE_DEV( dT ); } fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } opts.cleanup(); TESTING_FINALIZE(); return status; }
/* //////////////////////////////////////////////////////////////////////////// -- Testing sgegqr */ int main( int argc, char** argv) { TESTING_INIT(); real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time; float e1, e2, e3, e4, e5, *work; float c_neg_one = MAGMA_S_NEG_ONE; float c_one = MAGMA_S_ONE; float c_zero = MAGMA_S_ZERO; float *h_A, *h_R, *tau, *dtau, *h_work, *h_rwork, tmp[1]; float *d_A, *dwork; magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn; magma_int_t ione = 1, ldwork; magma_int_t ISEED[4] = {0,0,0,1}; magma_int_t status = 0; magma_opts opts; parse_opts( argc, argv, &opts ); opts.lapack |= opts.check; // check (-c) implies lapack (-l) // versions 1...4 are valid if (opts.version < 1 || opts.version > 4) { printf("Unknown version %d; exiting\n", opts.version ); return -1; } float tol, eps = lapackf77_slamch("E"); tol = 10* opts.tolerance * eps; printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||I-Q'Q||_F / M ||I-Q'Q||_I / M ||A-Q R||_I\n"); printf(" MAGMA / LAPACK MAGMA / LAPACK\n"); printf("==========================================================================================================\n"); for( int itest = 0; itest < opts.ntest; ++itest ) { for( int iter = 0; iter < opts.niter; ++iter ) { M = opts.msize[itest]; N = opts.nsize[itest]; if (N > 128) { printf("%5d %5d skipping because sgegqr requires N <= 128\n", (int) M, (int) N); continue; } if (M < N) { printf("%5d %5d skipping because sgegqr requires M >= N\n", (int) M, (int) N); continue; } min_mn = min(M, N); lda = M; n2 = lda*N; ldda = ((M+31)/32)*32; gflops = FLOPS_SGEQRF( M, N ) / 1e9 + FLOPS_SORGQR( M, N, N ) / 1e9; // query for workspace size lwork = -1; lapackf77_sgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info); lwork = (magma_int_t)MAGMA_S_REAL( tmp[0] ); lwork = max(lwork, 3*N*N); ldwork = N*N; if (opts.version == 2) { ldwork = 3*N*N + min_mn; } TESTING_MALLOC_PIN( tau, float, min_mn ); TESTING_MALLOC_PIN( h_work, float, lwork ); TESTING_MALLOC_PIN(h_rwork, float, lwork ); TESTING_MALLOC_CPU( h_A, float, n2 ); TESTING_MALLOC_CPU( h_R, float, n2 ); TESTING_MALLOC_CPU( work, float, M ); TESTING_MALLOC_DEV( d_A, float, ldda*N ); TESTING_MALLOC_DEV( dtau, float, min_mn ); TESTING_MALLOC_DEV( dwork, float, ldwork ); /* Initialize the matrix */ lapackf77_slarnv( &ione, ISEED, &n2, h_A ); lapackf77_slacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda ); magma_ssetmatrix( M, N, h_R, lda, d_A, ldda ); // warmup magma_sgegqr_gpu( 1, M, N, d_A, ldda, dwork, h_work, &info ); magma_ssetmatrix( M, N, h_R, lda, d_A, ldda ); /* ==================================================================== Performs operation using MAGMA =================================================================== */ gpu_time = magma_sync_wtime( 0 ); magma_sgegqr_gpu( opts.version, M, N, d_A, ldda, dwork, h_rwork, &info ); gpu_time = magma_sync_wtime( 0 ) - gpu_time; gpu_perf = gflops / gpu_time; if (info != 0) printf("magma_sgegqr returned error %d: %s.\n", (int) info, magma_strerror( info )); magma_sgetmatrix( M, N, d_A, ldda, h_R, M ); // Regenerate R // blasf77_sgemm("t", "n", &N, &N, &M, &c_one, h_R, &M, h_A, &M, &c_zero, h_rwork, &N); // magma_sprint(N, N, h_work, N); blasf77_strmm("r", "u", "n", "n", &M, &N, &c_one, h_rwork, &N, h_R, &M); blasf77_saxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione ); e5 = lapackf77_slange("i", &M, &N, h_R, &M, work) / lapackf77_slange("i", &M, &N, h_A, &lda, work); magma_sgetmatrix( M, N, d_A, ldda, h_R, M ); if ( opts.lapack ) { /* ===================================================================== Performs operation using LAPACK =================================================================== */ cpu_time = magma_wtime(); /* Orthogonalize on the CPU */ lapackf77_sgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info); lapackf77_sorgqr(&M, &N, &N, h_A, &lda, tau, h_work, &lwork, &info ); cpu_time = magma_wtime() - cpu_time; cpu_perf = gflops / cpu_time; if (info != 0) printf("lapackf77_sorgqr returned error %d: %s.\n", (int) info, magma_strerror( info )); /* ===================================================================== Check the result compared to LAPACK =================================================================== */ blasf77_sgemm("t", "n", &N, &N, &M, &c_one, h_R, &M, h_R, &M, &c_zero, h_work, &N); for(int ii = 0; ii < N*N; ii += N+1 ) { h_work[ii] = MAGMA_S_SUB(h_work[ii], c_one); } e1 = lapackf77_slange("f", &N, &N, h_work, &N, work) / N; e3 = lapackf77_slange("i", &N, &N, h_work, &N, work) / N; blasf77_sgemm("t", "n", &N, &N, &M, &c_one, h_A, &M, h_A, &M, &c_zero, h_work, &N); for(int ii = 0; ii < N*N; ii += N+1 ) { h_work[ii] = MAGMA_S_SUB(h_work[ii], c_one); } e2 = lapackf77_slange("f", &N, &N, h_work, &N, work) / N; e4 = lapackf77_slange("i", &N, &N, h_work, &N, work) / N; printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e / %8.2e %8.2e / %8.2e %8.2e %s\n", (int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time, e1, e2, e3, e4, e5, (e1 < tol ? "ok" : "failed")); status += ! (e1 < tol); } else { printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n", (int) M, (int) N, gpu_perf, 1000.*gpu_time ); } TESTING_FREE_PIN( tau ); TESTING_FREE_PIN( h_work ); TESTING_FREE_PIN( h_rwork ); TESTING_FREE_CPU( h_A ); TESTING_FREE_CPU( h_R ); TESTING_FREE_CPU( work ); TESTING_FREE_DEV( d_A ); TESTING_FREE_DEV( dtau ); TESTING_FREE_DEV( dwork ); fflush( stdout ); } if ( opts.niter > 1 ) { printf( "\n" ); } } TESTING_FINALIZE(); return status; }
/** Purpose ------- SORGQR generates an M-by-N REAL matrix Q with orthonormal columns, which is defined as the first N columns of a product of K elementary reflectors of order M Q = H(1) H(2) . . . H(k) as returned by SGEQRF_GPU. Arguments --------- @param[in] m INTEGER The number of rows of the matrix Q. M >= 0. @param[in] n INTEGER The number of columns of the matrix Q. M >= N >= 0. @param[in] k INTEGER The number of elementary reflectors whose product defines the matrix Q. N >= K >= 0. @param[in,out] dA REAL array A on the GPU, dimension (LDDA,N). On entry, the i-th column must contain the vector which defines the elementary reflector H(i), for i = 1,2,...,k, as returned by SGEQRF_GPU in the first k columns of its array argument A. On exit, the M-by-N matrix Q. @param[in] ldda INTEGER The first dimension of the array A. LDDA >= max(1,M). @param[in] tau REAL array, dimension (K) TAU(i) must contain the scalar factor of the elementary reflector H(i), as returned by SGEQRF_GPU. @param[in] dT (workspace) REAL work space array on the GPU, dimension (2*MIN(M, N) + (N+31)/32*32 )*NB. This must be the 6th argument of magma_sgeqrf_gpu [ note that if N here is bigger than N in magma_sgeqrf_gpu, the workspace requirement DT in magma_sgeqrf_gpu must be as specified in this routine ]. @param[in] nb INTEGER This is the block size used in SGEQRF_GPU, and correspondingly the size of the T matrices, used in the factorization, and stored in DT. @param[out] info INTEGER - = 0: successful exit - < 0: if INFO = -i, the i-th argument has an illegal value @ingroup magma_sgeqrf_comp ********************************************************************/ extern "C" magma_int_t magma_sorgqr_gpu(magma_int_t m, magma_int_t n, magma_int_t k, float *dA, magma_int_t ldda, float *tau, float *dT, magma_int_t nb, magma_int_t *info) { #define dA(i,j) (dA + (i) + (j)*ldda) #define dT(j) (dT + (j)*nb) float c_zero = MAGMA_S_ZERO; float c_one = MAGMA_S_ONE; magma_int_t m_kk, n_kk, k_kk, mi; magma_int_t lwork, lpanel; magma_int_t i, ib, ki, kk, iinfo; magma_int_t lddwork; float *dV, *dW; float *work, *panel; *info = 0; if (m < 0) { *info = -1; } else if ((n < 0) || (n > m)) { *info = -2; } else if ((k < 0) || (k > n)) { *info = -3; } else if (ldda < max(1,m)) { *info = -5; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } if (n <= 0) { return *info; } // first kk columns are handled by blocked method. // ki is start of 2nd-to-last block if ((nb > 1) && (nb < k)) { ki = (k - nb - 1) / nb * nb; kk = min( k, ki+nb ); } else { ki = 0; kk = 0; } // Allocate CPU work space // n*nb for sorgqr workspace // (m - kk)*(n - kk) for last block's panel lwork = n*nb; lpanel = (m - kk)*(n - kk); magma_smalloc_cpu( &work, lwork + lpanel ); if ( work == NULL ) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } panel = work + lwork; // Allocate work space on GPU if (MAGMA_SUCCESS != magma_smalloc( &dV, ldda*nb )) { magma_free_cpu( work ); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } // dT workspace has: // 2*min(m,n)*nb for T and R^{-1} matrices from geqrf // ((n+31)/32*32 )*nb for dW larfb workspace. lddwork = min(m,n); dW = dT + 2*lddwork*nb; magma_queue_t stream; magma_queue_create( &stream ); // Use unblocked code for the last or only block. if (kk < n) { m_kk = m - kk; n_kk = n - kk; k_kk = k - kk; magma_sgetmatrix( m_kk, k_kk, dA(kk, kk), ldda, panel, m_kk ); lapackf77_sorgqr( &m_kk, &n_kk, &k_kk, panel, &m_kk, &tau[kk], work, &lwork, &iinfo ); magma_ssetmatrix( m_kk, n_kk, panel, m_kk, dA(kk, kk), ldda ); // Set A(1:kk,kk+1:n) to zero. magmablas_slaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(0, kk), ldda ); } if (kk > 0) { // Use blocked code // stream: copy Aii to V --> laset --> laset --> larfb --> [next] // CPU has no computation magmablasSetKernelStream( stream ); for (i = ki; i >= 0; i -= nb) { ib = min( nb, k-i ); mi = m - i; // Copy current panel on the GPU from dA to dV magma_scopymatrix_async( mi, ib, dA(i,i), ldda, dV, ldda, stream ); // set panel to identity magmablas_slaset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda ); magmablas_slaset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda ); if (i < n) { // Apply H to A(i:m,i:n) from the left magma_slarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, mi, n-i, ib, dV, ldda, dT(i), nb, dA(i, i), ldda, dW, lddwork ); } } } magma_queue_sync( stream ); magmablasSetKernelStream( NULL ); magma_free( dV ); magma_free_cpu( work ); magma_queue_destroy( stream ); return *info; } /* magma_sorgqr_gpu */
extern "C" magma_int_t magma_sorgqr_gpu(magma_int_t m, magma_int_t n, magma_int_t k, float *dA, magma_int_t ldda, float *tau, float *dT, magma_int_t nb, magma_int_t *info) { /* -- MAGMA (version 1.3.0) -- Univ. of Tennessee, Knoxville Univ. of California, Berkeley Univ. of Colorado, Denver November 2012 Purpose ======= SORGQR generates an M-by-N REAL matrix Q with orthonormal columns, which is defined as the first N columns of a product of K elementary reflectors of order M Q = H(1) H(2) . . . H(k) as returned by SGEQRF_GPU. Arguments ========= M (input) INTEGER The number of rows of the matrix Q. M >= 0. N (input) INTEGER The number of columns of the matrix Q. M >= N >= 0. K (input) INTEGER The number of elementary reflectors whose product defines the matrix Q. N >= K >= 0. DA (input/output) REAL array A on the GPU, dimension (LDDA,N). On entry, the i-th column must contain the vector which defines the elementary reflector H(i), for i = 1,2,...,k, as returned by SGEQRF_GPU in the first k columns of its array argument A. On exit, the M-by-N matrix Q. LDDA (input) INTEGER The first dimension of the array A. LDDA >= max(1,M). TAU (input) REAL array, dimension (K) TAU(i) must contain the scalar factor of the elementary reflector H(i), as returned by SGEQRF_GPU. DT (input/workspace) REAL work space array on the GPU, dimension (2*MIN(M, N) + (N+31)/32*32 )*NB. This must be the 6th argument of magma_sgeqrf_gpu [ note that if N here is bigger than N in magma_sgeqrf_gpu, the workspace requirement DT in magma_sgeqrf_gpu must be as specified in this routine ]. NB (input) INTEGER This is the block size used in SGEQRF_GPU, and correspondingly the size of the T matrices, used in the factorization, and stored in DT. INFO (output) INTEGER = 0: successful exit < 0: if INFO = -i, the i-th argument has an illegal value ===================================================================== */ #define dA(i,j) (dA + (i) + (j)*ldda) #define dT(j) (dT + (j)*nb) magma_int_t m_kk, n_kk, k_kk, mi; magma_int_t lwork, lpanel; magma_int_t i, ib, ki, kk, iinfo; magma_int_t lddwork; float *dV, *dW; float *work, *panel; *info = 0; if (m < 0) { *info = -1; } else if ((n < 0) || (n > m)) { *info = -2; } else if ((k < 0) || (k > n)) { *info = -3; } else if (ldda < max(1,m)) { *info = -5; } if (*info != 0) { magma_xerbla( __func__, -(*info) ); return *info; } if (n <= 0) { return *info; } // first kk columns are handled by blocked method. if ((nb > 1) && (nb < k)) { ki = (k - nb - 1) / nb * nb; kk = min( k, ki+nb ); } else { kk = 0; } // Allocate CPU work space // n*nb for sorgqr workspace // (m - kk)*(n - kk) for last block's panel lwork = n*nb; lpanel = (m - kk)*(n - kk); magma_smalloc_cpu( &work, lwork + lpanel ); if ( work == NULL ) { *info = MAGMA_ERR_HOST_ALLOC; return *info; } panel = work + lwork; // Allocate work space on GPU if (MAGMA_SUCCESS != magma_smalloc( &dV, ldda*nb )) { magma_free_cpu( work ); *info = MAGMA_ERR_DEVICE_ALLOC; return *info; } // dT workspace has: // 2*min(m,n)*nb for T and R^{-1} matrices from geqrf // ((n+31)/32*32 )*nb for dW larfb workspace. lddwork = min(m,n); dW = dT + 2*lddwork*nb; cudaStream_t stream; magma_queue_create( &stream ); // Use unblocked code for the last or only block. if (kk < n) { m_kk = m - kk; n_kk = n - kk; k_kk = k - kk; magma_sgetmatrix( m_kk, n_kk, dA(kk, kk), ldda, panel, m_kk ); lapackf77_sorgqr( &m_kk, &n_kk, &k_kk, panel, &m_kk, &tau[kk], work, &lwork, &iinfo ); magma_ssetmatrix( m_kk, n_kk, panel, m_kk, dA(kk, kk), ldda ); // Set A(1:kk,kk+1:n) to zero. magmablas_slaset( MagmaUpperLower, kk, n - kk, dA(0, kk), ldda ); } if (kk > 0) { // Use blocked code // stream: copy Aii to V --> laset --> laset --> larfb --> [next] // CPU has no computation magmablasSetKernelStream( stream ); for (i = ki; i >= 0; i -= nb) { ib = min( nb, k-i ); mi = m - i; // Copy current panel on the GPU from dA to dV magma_scopymatrix_async( mi, ib, dA(i,i), ldda, dV, ldda, stream ); // set panel to identity magmablas_slaset( MagmaUpperLower, i, ib, dA(0, i), ldda ); magmablas_slaset_identity( mi, ib, dA(i, i), ldda ); if (i < n) { // Apply H to A(i:m,i:n) from the left magma_slarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, mi, n-i, ib, dV, ldda, dT(i), nb, dA(i, i), ldda, dW, lddwork ); } } } magma_queue_sync( stream ); magmablasSetKernelStream( NULL ); magma_free( dV ); magma_free_cpu( work ); magma_queue_destroy( stream ); return *info; } /* magma_sorgqr_gpu */